2 * Author......: Jens Steube <jens.steube@gmail.com>
8 #include "include/constants.h"
9 #include "include/kernel_vendor.h"
28 #include "include/kernel_functions.c"
29 #include "types_amd.c"
30 #include "common_amd.c"
31 #include "include/rp_gpu.h"
35 #define VECT_COMPARE_S "check_single_vect1_comp4.c"
36 #define VECT_COMPARE_M "check_multi_vect1_comp4.c"
40 #define VECT_COMPARE_S "check_single_vect2_comp4.c"
41 #define VECT_COMPARE_M "check_multi_vect2_comp4.c"
45 #define VECT_COMPARE_S "check_single_vect4_comp4.c"
46 #define VECT_COMPARE_M "check_multi_vect4_comp4.c"
49 #define PERM_OP(a,b,tt,n,m) \
59 #define HPERM_OP(a,tt,n,m) \
65 tt = tt >> (16 + n); \
69 __constant u32 c_SPtrans[8][64] =
72 0x02080800, 0x00080000, 0x02000002, 0x02080802,
73 0x02000000, 0x00080802, 0x00080002, 0x02000002,
74 0x00080802, 0x02080800, 0x02080000, 0x00000802,
75 0x02000802, 0x02000000, 0x00000000, 0x00080002,
76 0x00080000, 0x00000002, 0x02000800, 0x00080800,
77 0x02080802, 0x02080000, 0x00000802, 0x02000800,
78 0x00000002, 0x00000800, 0x00080800, 0x02080002,
79 0x00000800, 0x02000802, 0x02080002, 0x00000000,
80 0x00000000, 0x02080802, 0x02000800, 0x00080002,
81 0x02080800, 0x00080000, 0x00000802, 0x02000800,
82 0x02080002, 0x00000800, 0x00080800, 0x02000002,
83 0x00080802, 0x00000002, 0x02000002, 0x02080000,
84 0x02080802, 0x00080800, 0x02080000, 0x02000802,
85 0x02000000, 0x00000802, 0x00080002, 0x00000000,
86 0x00080000, 0x02000000, 0x02000802, 0x02080800,
87 0x00000002, 0x02080002, 0x00000800, 0x00080802,
89 0x40108010, 0x00000000, 0x00108000, 0x40100000,
90 0x40000010, 0x00008010, 0x40008000, 0x00108000,
91 0x00008000, 0x40100010, 0x00000010, 0x40008000,
92 0x00100010, 0x40108000, 0x40100000, 0x00000010,
93 0x00100000, 0x40008010, 0x40100010, 0x00008000,
94 0x00108010, 0x40000000, 0x00000000, 0x00100010,
95 0x40008010, 0x00108010, 0x40108000, 0x40000010,
96 0x40000000, 0x00100000, 0x00008010, 0x40108010,
97 0x00100010, 0x40108000, 0x40008000, 0x00108010,
98 0x40108010, 0x00100010, 0x40000010, 0x00000000,
99 0x40000000, 0x00008010, 0x00100000, 0x40100010,
100 0x00008000, 0x40000000, 0x00108010, 0x40008010,
101 0x40108000, 0x00008000, 0x00000000, 0x40000010,
102 0x00000010, 0x40108010, 0x00108000, 0x40100000,
103 0x40100010, 0x00100000, 0x00008010, 0x40008000,
104 0x40008010, 0x00000010, 0x40100000, 0x00108000,
106 0x04000001, 0x04040100, 0x00000100, 0x04000101,
107 0x00040001, 0x04000000, 0x04000101, 0x00040100,
108 0x04000100, 0x00040000, 0x04040000, 0x00000001,
109 0x04040101, 0x00000101, 0x00000001, 0x04040001,
110 0x00000000, 0x00040001, 0x04040100, 0x00000100,
111 0x00000101, 0x04040101, 0x00040000, 0x04000001,
112 0x04040001, 0x04000100, 0x00040101, 0x04040000,
113 0x00040100, 0x00000000, 0x04000000, 0x00040101,
114 0x04040100, 0x00000100, 0x00000001, 0x00040000,
115 0x00000101, 0x00040001, 0x04040000, 0x04000101,
116 0x00000000, 0x04040100, 0x00040100, 0x04040001,
117 0x00040001, 0x04000000, 0x04040101, 0x00000001,
118 0x00040101, 0x04000001, 0x04000000, 0x04040101,
119 0x00040000, 0x04000100, 0x04000101, 0x00040100,
120 0x04000100, 0x00000000, 0x04040001, 0x00000101,
121 0x04000001, 0x00040101, 0x00000100, 0x04040000,
123 0x00401008, 0x10001000, 0x00000008, 0x10401008,
124 0x00000000, 0x10400000, 0x10001008, 0x00400008,
125 0x10401000, 0x10000008, 0x10000000, 0x00001008,
126 0x10000008, 0x00401008, 0x00400000, 0x10000000,
127 0x10400008, 0x00401000, 0x00001000, 0x00000008,
128 0x00401000, 0x10001008, 0x10400000, 0x00001000,
129 0x00001008, 0x00000000, 0x00400008, 0x10401000,
130 0x10001000, 0x10400008, 0x10401008, 0x00400000,
131 0x10400008, 0x00001008, 0x00400000, 0x10000008,
132 0x00401000, 0x10001000, 0x00000008, 0x10400000,
133 0x10001008, 0x00000000, 0x00001000, 0x00400008,
134 0x00000000, 0x10400008, 0x10401000, 0x00001000,
135 0x10000000, 0x10401008, 0x00401008, 0x00400000,
136 0x10401008, 0x00000008, 0x10001000, 0x00401008,
137 0x00400008, 0x00401000, 0x10400000, 0x10001008,
138 0x00001008, 0x10000000, 0x10000008, 0x10401000,
140 0x08000000, 0x00010000, 0x00000400, 0x08010420,
141 0x08010020, 0x08000400, 0x00010420, 0x08010000,
142 0x00010000, 0x00000020, 0x08000020, 0x00010400,
143 0x08000420, 0x08010020, 0x08010400, 0x00000000,
144 0x00010400, 0x08000000, 0x00010020, 0x00000420,
145 0x08000400, 0x00010420, 0x00000000, 0x08000020,
146 0x00000020, 0x08000420, 0x08010420, 0x00010020,
147 0x08010000, 0x00000400, 0x00000420, 0x08010400,
148 0x08010400, 0x08000420, 0x00010020, 0x08010000,
149 0x00010000, 0x00000020, 0x08000020, 0x08000400,
150 0x08000000, 0x00010400, 0x08010420, 0x00000000,
151 0x00010420, 0x08000000, 0x00000400, 0x00010020,
152 0x08000420, 0x00000400, 0x00000000, 0x08010420,
153 0x08010020, 0x08010400, 0x00000420, 0x00010000,
154 0x00010400, 0x08010020, 0x08000400, 0x00000420,
155 0x00000020, 0x00010420, 0x08010000, 0x08000020,
157 0x80000040, 0x00200040, 0x00000000, 0x80202000,
158 0x00200040, 0x00002000, 0x80002040, 0x00200000,
159 0x00002040, 0x80202040, 0x00202000, 0x80000000,
160 0x80002000, 0x80000040, 0x80200000, 0x00202040,
161 0x00200000, 0x80002040, 0x80200040, 0x00000000,
162 0x00002000, 0x00000040, 0x80202000, 0x80200040,
163 0x80202040, 0x80200000, 0x80000000, 0x00002040,
164 0x00000040, 0x00202000, 0x00202040, 0x80002000,
165 0x00002040, 0x80000000, 0x80002000, 0x00202040,
166 0x80202000, 0x00200040, 0x00000000, 0x80002000,
167 0x80000000, 0x00002000, 0x80200040, 0x00200000,
168 0x00200040, 0x80202040, 0x00202000, 0x00000040,
169 0x80202040, 0x00202000, 0x00200000, 0x80002040,
170 0x80000040, 0x80200000, 0x00202040, 0x00000000,
171 0x00002000, 0x80000040, 0x80002040, 0x80202000,
172 0x80200000, 0x00002040, 0x00000040, 0x80200040,
174 0x00004000, 0x00000200, 0x01000200, 0x01000004,
175 0x01004204, 0x00004004, 0x00004200, 0x00000000,
176 0x01000000, 0x01000204, 0x00000204, 0x01004000,
177 0x00000004, 0x01004200, 0x01004000, 0x00000204,
178 0x01000204, 0x00004000, 0x00004004, 0x01004204,
179 0x00000000, 0x01000200, 0x01000004, 0x00004200,
180 0x01004004, 0x00004204, 0x01004200, 0x00000004,
181 0x00004204, 0x01004004, 0x00000200, 0x01000000,
182 0x00004204, 0x01004000, 0x01004004, 0x00000204,
183 0x00004000, 0x00000200, 0x01000000, 0x01004004,
184 0x01000204, 0x00004204, 0x00004200, 0x00000000,
185 0x00000200, 0x01000004, 0x00000004, 0x01000200,
186 0x00000000, 0x01000204, 0x01000200, 0x00004200,
187 0x00000204, 0x00004000, 0x01004204, 0x01000000,
188 0x01004200, 0x00000004, 0x00004004, 0x01004204,
189 0x01000004, 0x01004200, 0x01004000, 0x00004004,
191 0x20800080, 0x20820000, 0x00020080, 0x00000000,
192 0x20020000, 0x00800080, 0x20800000, 0x20820080,
193 0x00000080, 0x20000000, 0x00820000, 0x00020080,
194 0x00820080, 0x20020080, 0x20000080, 0x20800000,
195 0x00020000, 0x00820080, 0x00800080, 0x20020000,
196 0x20820080, 0x20000080, 0x00000000, 0x00820000,
197 0x20000000, 0x00800000, 0x20020080, 0x20800080,
198 0x00800000, 0x00020000, 0x20820000, 0x00000080,
199 0x00800000, 0x00020000, 0x20000080, 0x20820080,
200 0x00020080, 0x20000000, 0x00000000, 0x00820000,
201 0x20800080, 0x20020080, 0x20020000, 0x00800080,
202 0x20820000, 0x00000080, 0x00800080, 0x20020000,
203 0x20820080, 0x00800000, 0x20800000, 0x20000080,
204 0x00820000, 0x00020080, 0x20020080, 0x20800000,
205 0x00000080, 0x20820000, 0x00820080, 0x00000000,
206 0x20000000, 0x20800080, 0x00020000, 0x00820080,
209 __constant u32 c_skb[8][64] =
211 /* for C bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
212 0x00000000, 0x00000010, 0x20000000, 0x20000010,
213 0x00010000, 0x00010010, 0x20010000, 0x20010010,
214 0x00000800, 0x00000810, 0x20000800, 0x20000810,
215 0x00010800, 0x00010810, 0x20010800, 0x20010810,
216 0x00000020, 0x00000030, 0x20000020, 0x20000030,
217 0x00010020, 0x00010030, 0x20010020, 0x20010030,
218 0x00000820, 0x00000830, 0x20000820, 0x20000830,
219 0x00010820, 0x00010830, 0x20010820, 0x20010830,
220 0x00080000, 0x00080010, 0x20080000, 0x20080010,
221 0x00090000, 0x00090010, 0x20090000, 0x20090010,
222 0x00080800, 0x00080810, 0x20080800, 0x20080810,
223 0x00090800, 0x00090810, 0x20090800, 0x20090810,
224 0x00080020, 0x00080030, 0x20080020, 0x20080030,
225 0x00090020, 0x00090030, 0x20090020, 0x20090030,
226 0x00080820, 0x00080830, 0x20080820, 0x20080830,
227 0x00090820, 0x00090830, 0x20090820, 0x20090830,
228 /* for C bits (numbered as per FIPS 46) 7 8 10 11 12 13 */
229 0x00000000, 0x02000000, 0x00002000, 0x02002000,
230 0x00200000, 0x02200000, 0x00202000, 0x02202000,
231 0x00000004, 0x02000004, 0x00002004, 0x02002004,
232 0x00200004, 0x02200004, 0x00202004, 0x02202004,
233 0x00000400, 0x02000400, 0x00002400, 0x02002400,
234 0x00200400, 0x02200400, 0x00202400, 0x02202400,
235 0x00000404, 0x02000404, 0x00002404, 0x02002404,
236 0x00200404, 0x02200404, 0x00202404, 0x02202404,
237 0x10000000, 0x12000000, 0x10002000, 0x12002000,
238 0x10200000, 0x12200000, 0x10202000, 0x12202000,
239 0x10000004, 0x12000004, 0x10002004, 0x12002004,
240 0x10200004, 0x12200004, 0x10202004, 0x12202004,
241 0x10000400, 0x12000400, 0x10002400, 0x12002400,
242 0x10200400, 0x12200400, 0x10202400, 0x12202400,
243 0x10000404, 0x12000404, 0x10002404, 0x12002404,
244 0x10200404, 0x12200404, 0x10202404, 0x12202404,
245 /* for C bits (numbered as per FIPS 46) 14 15 16 17 19 20 */
246 0x00000000, 0x00000001, 0x00040000, 0x00040001,
247 0x01000000, 0x01000001, 0x01040000, 0x01040001,
248 0x00000002, 0x00000003, 0x00040002, 0x00040003,
249 0x01000002, 0x01000003, 0x01040002, 0x01040003,
250 0x00000200, 0x00000201, 0x00040200, 0x00040201,
251 0x01000200, 0x01000201, 0x01040200, 0x01040201,
252 0x00000202, 0x00000203, 0x00040202, 0x00040203,
253 0x01000202, 0x01000203, 0x01040202, 0x01040203,
254 0x08000000, 0x08000001, 0x08040000, 0x08040001,
255 0x09000000, 0x09000001, 0x09040000, 0x09040001,
256 0x08000002, 0x08000003, 0x08040002, 0x08040003,
257 0x09000002, 0x09000003, 0x09040002, 0x09040003,
258 0x08000200, 0x08000201, 0x08040200, 0x08040201,
259 0x09000200, 0x09000201, 0x09040200, 0x09040201,
260 0x08000202, 0x08000203, 0x08040202, 0x08040203,
261 0x09000202, 0x09000203, 0x09040202, 0x09040203,
262 /* for C bits (numbered as per FIPS 46) 21 23 24 26 27 28 */
263 0x00000000, 0x00100000, 0x00000100, 0x00100100,
264 0x00000008, 0x00100008, 0x00000108, 0x00100108,
265 0x00001000, 0x00101000, 0x00001100, 0x00101100,
266 0x00001008, 0x00101008, 0x00001108, 0x00101108,
267 0x04000000, 0x04100000, 0x04000100, 0x04100100,
268 0x04000008, 0x04100008, 0x04000108, 0x04100108,
269 0x04001000, 0x04101000, 0x04001100, 0x04101100,
270 0x04001008, 0x04101008, 0x04001108, 0x04101108,
271 0x00020000, 0x00120000, 0x00020100, 0x00120100,
272 0x00020008, 0x00120008, 0x00020108, 0x00120108,
273 0x00021000, 0x00121000, 0x00021100, 0x00121100,
274 0x00021008, 0x00121008, 0x00021108, 0x00121108,
275 0x04020000, 0x04120000, 0x04020100, 0x04120100,
276 0x04020008, 0x04120008, 0x04020108, 0x04120108,
277 0x04021000, 0x04121000, 0x04021100, 0x04121100,
278 0x04021008, 0x04121008, 0x04021108, 0x04121108,
279 /* for D bits (numbered as per FIPS 46) 1 2 3 4 5 6 */
280 0x00000000, 0x10000000, 0x00010000, 0x10010000,
281 0x00000004, 0x10000004, 0x00010004, 0x10010004,
282 0x20000000, 0x30000000, 0x20010000, 0x30010000,
283 0x20000004, 0x30000004, 0x20010004, 0x30010004,
284 0x00100000, 0x10100000, 0x00110000, 0x10110000,
285 0x00100004, 0x10100004, 0x00110004, 0x10110004,
286 0x20100000, 0x30100000, 0x20110000, 0x30110000,
287 0x20100004, 0x30100004, 0x20110004, 0x30110004,
288 0x00001000, 0x10001000, 0x00011000, 0x10011000,
289 0x00001004, 0x10001004, 0x00011004, 0x10011004,
290 0x20001000, 0x30001000, 0x20011000, 0x30011000,
291 0x20001004, 0x30001004, 0x20011004, 0x30011004,
292 0x00101000, 0x10101000, 0x00111000, 0x10111000,
293 0x00101004, 0x10101004, 0x00111004, 0x10111004,
294 0x20101000, 0x30101000, 0x20111000, 0x30111000,
295 0x20101004, 0x30101004, 0x20111004, 0x30111004,
296 /* for D bits (numbered as per FIPS 46) 8 9 11 12 13 14 */
297 0x00000000, 0x08000000, 0x00000008, 0x08000008,
298 0x00000400, 0x08000400, 0x00000408, 0x08000408,
299 0x00020000, 0x08020000, 0x00020008, 0x08020008,
300 0x00020400, 0x08020400, 0x00020408, 0x08020408,
301 0x00000001, 0x08000001, 0x00000009, 0x08000009,
302 0x00000401, 0x08000401, 0x00000409, 0x08000409,
303 0x00020001, 0x08020001, 0x00020009, 0x08020009,
304 0x00020401, 0x08020401, 0x00020409, 0x08020409,
305 0x02000000, 0x0A000000, 0x02000008, 0x0A000008,
306 0x02000400, 0x0A000400, 0x02000408, 0x0A000408,
307 0x02020000, 0x0A020000, 0x02020008, 0x0A020008,
308 0x02020400, 0x0A020400, 0x02020408, 0x0A020408,
309 0x02000001, 0x0A000001, 0x02000009, 0x0A000009,
310 0x02000401, 0x0A000401, 0x02000409, 0x0A000409,
311 0x02020001, 0x0A020001, 0x02020009, 0x0A020009,
312 0x02020401, 0x0A020401, 0x02020409, 0x0A020409,
313 /* for D bits (numbered as per FIPS 46) 16 17 18 19 20 21 */
314 0x00000000, 0x00000100, 0x00080000, 0x00080100,
315 0x01000000, 0x01000100, 0x01080000, 0x01080100,
316 0x00000010, 0x00000110, 0x00080010, 0x00080110,
317 0x01000010, 0x01000110, 0x01080010, 0x01080110,
318 0x00200000, 0x00200100, 0x00280000, 0x00280100,
319 0x01200000, 0x01200100, 0x01280000, 0x01280100,
320 0x00200010, 0x00200110, 0x00280010, 0x00280110,
321 0x01200010, 0x01200110, 0x01280010, 0x01280110,
322 0x00000200, 0x00000300, 0x00080200, 0x00080300,
323 0x01000200, 0x01000300, 0x01080200, 0x01080300,
324 0x00000210, 0x00000310, 0x00080210, 0x00080310,
325 0x01000210, 0x01000310, 0x01080210, 0x01080310,
326 0x00200200, 0x00200300, 0x00280200, 0x00280300,
327 0x01200200, 0x01200300, 0x01280200, 0x01280300,
328 0x00200210, 0x00200310, 0x00280210, 0x00280310,
329 0x01200210, 0x01200310, 0x01280210, 0x01280310,
330 /* for D bits (numbered as per FIPS 46) 22 23 24 25 27 28 */
331 0x00000000, 0x04000000, 0x00040000, 0x04040000,
332 0x00000002, 0x04000002, 0x00040002, 0x04040002,
333 0x00002000, 0x04002000, 0x00042000, 0x04042000,
334 0x00002002, 0x04002002, 0x00042002, 0x04042002,
335 0x00000020, 0x04000020, 0x00040020, 0x04040020,
336 0x00000022, 0x04000022, 0x00040022, 0x04040022,
337 0x00002020, 0x04002020, 0x00042020, 0x04042020,
338 0x00002022, 0x04002022, 0x00042022, 0x04042022,
339 0x00000800, 0x04000800, 0x00040800, 0x04040800,
340 0x00000802, 0x04000802, 0x00040802, 0x04040802,
341 0x00002800, 0x04002800, 0x00042800, 0x04042800,
342 0x00002802, 0x04002802, 0x00042802, 0x04042802,
343 0x00000820, 0x04000820, 0x00040820, 0x04040820,
344 0x00000822, 0x04000822, 0x00040822, 0x04040822,
345 0x00002820, 0x04002820, 0x00042820, 0x04042820,
346 0x00002822, 0x04002822, 0x00042822, 0x04042822
349 __constant u32 shifts3s0[16] = { 1, 1, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 1 };
350 __constant u32 shifts3s1[16] = { 27, 27, 26, 26, 26, 26, 26, 26, 27, 26, 26, 26, 26, 26, 26, 27 };
353 #define BOX(i,n,S) (u32x) ((S)[(n)][(i)])
357 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1])
361 #define BOX(i,n,S) (u32x) ((S)[(n)][(i).s0], (S)[(n)][(i).s1], (S)[(n)][(i).s2], (S)[(n)][(i).s3])
364 static void _des_crypt_encrypt (u32x iv[2], u32x data[2], u32x Kc[16], u32x Kd[16], __local u32 s_SPtrans[8][64])
370 for (u32 i = 0; i < 16; i += 2)
375 u = Kc[i + 0] ^ rotl32 (r, 30u);
376 t = Kd[i + 0] ^ rotl32 (r, 26u);
379 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
380 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
381 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
382 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
383 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
384 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
385 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
386 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
388 u = Kc[i + 1] ^ rotl32 (l, 30u);
389 t = Kd[i + 1] ^ rotl32 (l, 26u);
392 ^ BOX (amd_bfe (u, 0, 6), 0, s_SPtrans)
393 ^ BOX (amd_bfe (u, 8, 6), 2, s_SPtrans)
394 ^ BOX (amd_bfe (u, 16, 6), 4, s_SPtrans)
395 ^ BOX (amd_bfe (u, 24, 6), 6, s_SPtrans)
396 ^ BOX (amd_bfe (t, 0, 6), 1, s_SPtrans)
397 ^ BOX (amd_bfe (t, 8, 6), 3, s_SPtrans)
398 ^ BOX (amd_bfe (t, 16, 6), 5, s_SPtrans)
399 ^ BOX (amd_bfe (t, 24, 6), 7, s_SPtrans);
406 static void _des_crypt_keysetup (u32x c, u32x d, u32x Kc[16], u32x Kd[16], __local u32 s_skb[8][64])
410 PERM_OP (d, c, tt, 4, 0x0f0f0f0f);
411 HPERM_OP (c, tt, 2, 0xcccc0000);
412 HPERM_OP (d, tt, 2, 0xcccc0000);
413 PERM_OP (d, c, tt, 1, 0x55555555);
414 PERM_OP (c, d, tt, 8, 0x00ff00ff);
415 PERM_OP (d, c, tt, 1, 0x55555555);
417 d = ((d & 0x000000ff) << 16)
418 | ((d & 0x0000ff00) << 0)
419 | ((d & 0x00ff0000) >> 16)
420 | ((c & 0xf0000000) >> 4);
425 for (u32 i = 0; i < 16; i++)
427 c = c >> shifts3s0[i] | c << shifts3s1[i];
428 d = d >> shifts3s0[i] | d << shifts3s1[i];
433 const u32x c00 = (c >> 0) & 0x0000003f;
434 const u32x c06 = (c >> 6) & 0x00383003;
435 const u32x c07 = (c >> 7) & 0x0000003c;
436 const u32x c13 = (c >> 13) & 0x0000060f;
437 const u32x c20 = (c >> 20) & 0x00000001;
439 u32x s = BOX (((c00 >> 0) & 0xff), 0, s_skb)
440 | BOX (((c06 >> 0) & 0xff)
441 |((c07 >> 0) & 0xff), 1, s_skb)
442 | BOX (((c13 >> 0) & 0xff)
443 |((c06 >> 8) & 0xff), 2, s_skb)
444 | BOX (((c20 >> 0) & 0xff)
446 |((c06 >> 16) & 0xff), 3, s_skb);
448 const u32x d00 = (d >> 0) & 0x00003c3f;
449 const u32x d07 = (d >> 7) & 0x00003f03;
450 const u32x d21 = (d >> 21) & 0x0000000f;
451 const u32x d22 = (d >> 22) & 0x00000030;
453 u32x t = BOX (((d00 >> 0) & 0xff), 4, s_skb)
454 | BOX (((d07 >> 0) & 0xff)
455 |((d00 >> 8) & 0xff), 5, s_skb)
456 | BOX (((d07 >> 8) & 0xff), 6, s_skb)
457 | BOX (((d21 >> 0) & 0xff)
458 |((d22 >> 0) & 0xff), 7, s_skb);
460 Kc[i] = ((t << 16) | (s & 0x0000ffff));
461 Kd[i] = ((s >> 16) | (t & 0xffff0000));
465 static void transform_netntlmv1_key (const u32x w0, const u32x w1, u32x out[2])
468 const uchar4 t0 = as_uchar4 (w0);
469 const uchar4 t1 = as_uchar4 (w1);
474 k0.s0 = (t0.s0 >> 0);
475 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
476 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
477 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
478 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
479 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
480 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
481 k1.s3 = (t1.s2 << 1);
483 out[0] = as_uint (k0);
484 out[1] = as_uint (k1);
488 const uchar8 t0 = as_uchar8 (w0);
489 const uchar8 t1 = as_uchar8 (w1);
494 k0.s0 = (t0.s0 >> 0);
495 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
496 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
497 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
498 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
499 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
500 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
501 k1.s3 = (t1.s2 << 1);
503 k0.s4 = (t0.s4 >> 0);
504 k0.s5 = (t0.s4 << 7) | (t0.s5 >> 1);
505 k0.s6 = (t0.s5 << 6) | (t0.s6 >> 2);
506 k0.s7 = (t0.s6 << 5) | (t0.s7 >> 3);
507 k1.s4 = (t0.s7 << 4) | (t1.s4 >> 4);
508 k1.s5 = (t1.s4 << 3) | (t1.s5 >> 5);
509 k1.s6 = (t1.s5 << 2) | (t1.s6 >> 6);
510 k1.s7 = (t1.s6 << 1);
512 out[0] = as_uint2 (k0);
513 out[1] = as_uint2 (k1);
517 const uchar16 t0 = as_uchar16 (w0);
518 const uchar16 t1 = as_uchar16 (w1);
523 k0.s0 = (t0.s0 >> 0);
524 k0.s1 = (t0.s0 << 7) | (t0.s1 >> 1);
525 k0.s2 = (t0.s1 << 6) | (t0.s2 >> 2);
526 k0.s3 = (t0.s2 << 5) | (t0.s3 >> 3);
527 k1.s0 = (t0.s3 << 4) | (t1.s0 >> 4);
528 k1.s1 = (t1.s0 << 3) | (t1.s1 >> 5);
529 k1.s2 = (t1.s1 << 2) | (t1.s2 >> 6);
530 k1.s3 = (t1.s2 << 1);
532 k0.s4 = (t0.s4 >> 0);
533 k0.s5 = (t0.s4 << 7) | (t0.s5 >> 1);
534 k0.s6 = (t0.s5 << 6) | (t0.s6 >> 2);
535 k0.s7 = (t0.s6 << 5) | (t0.s7 >> 3);
536 k1.s4 = (t0.s7 << 4) | (t1.s4 >> 4);
537 k1.s5 = (t1.s4 << 3) | (t1.s5 >> 5);
538 k1.s6 = (t1.s5 << 2) | (t1.s6 >> 6);
539 k1.s7 = (t1.s6 << 1);
541 k0.s8 = (t0.s8 >> 0);
542 k0.s9 = (t0.s8 << 7) | (t0.s9 >> 1);
543 k0.sa = (t0.s9 << 6) | (t0.sa >> 2);
544 k0.sb = (t0.sa << 5) | (t0.sb >> 3);
545 k1.s8 = (t0.sb << 4) | (t1.s8 >> 4);
546 k1.s9 = (t1.s8 << 3) | (t1.s9 >> 5);
547 k1.sa = (t1.s9 << 2) | (t1.sa >> 6);
548 k1.sb = (t1.sa << 1);
550 k0.sc = (t0.sc >> 0);
551 k0.sd = (t0.sc << 7) | (t0.sd >> 1);
552 k0.se = (t0.sd << 6) | (t0.se >> 2);
553 k0.sf = (t0.se << 5) | (t0.sf >> 3);
554 k1.sc = (t0.sf << 4) | (t1.sc >> 4);
555 k1.sd = (t1.sc << 3) | (t1.sd >> 5);
556 k1.se = (t1.sd << 2) | (t1.se >> 6);
557 k1.sf = (t1.se << 1);
559 out[0] = as_uint4 (k0);
560 out[1] = as_uint4 (k1);
564 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_m04 (__global pw_t *pws, __global gpu_rule_t * rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
570 const u32 lid = get_local_id (0);
576 const u32 gid = get_global_id (0);
580 pw_buf0[0] = pws[gid].i[ 0];
581 pw_buf0[1] = pws[gid].i[ 1];
582 pw_buf0[2] = pws[gid].i[ 2];
583 pw_buf0[3] = pws[gid].i[ 3];
587 pw_buf1[0] = pws[gid].i[ 4];
588 pw_buf1[1] = pws[gid].i[ 5];
589 pw_buf1[2] = pws[gid].i[ 6];
590 pw_buf1[3] = pws[gid].i[ 7];
592 const u32 pw_len = pws[gid].pw_len;
598 __local u32 s_SPtrans[8][64];
599 __local u32 s_skb[8][64];
601 s_SPtrans[0][lid] = c_SPtrans[0][lid];
602 s_SPtrans[1][lid] = c_SPtrans[1][lid];
603 s_SPtrans[2][lid] = c_SPtrans[2][lid];
604 s_SPtrans[3][lid] = c_SPtrans[3][lid];
605 s_SPtrans[4][lid] = c_SPtrans[4][lid];
606 s_SPtrans[5][lid] = c_SPtrans[5][lid];
607 s_SPtrans[6][lid] = c_SPtrans[6][lid];
608 s_SPtrans[7][lid] = c_SPtrans[7][lid];
610 s_skb[0][lid] = c_skb[0][lid];
611 s_skb[1][lid] = c_skb[1][lid];
612 s_skb[2][lid] = c_skb[2][lid];
613 s_skb[3][lid] = c_skb[3][lid];
614 s_skb[4][lid] = c_skb[4][lid];
615 s_skb[5][lid] = c_skb[5][lid];
616 s_skb[6][lid] = c_skb[6][lid];
617 s_skb[7][lid] = c_skb[7][lid];
619 barrier (CLK_LOCAL_MEM_FENCE);
621 if (gid >= gid_max) return;
627 const u32 s0 = salt_bufs[salt_pos].salt_buf[0];
628 const u32 s1 = salt_bufs[salt_pos].salt_buf[1];
629 const u32 s2 = salt_bufs[salt_pos].salt_buf[2];
640 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
670 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
672 append_0x80_2 (w0, w1, out_len);
679 make_unicode (w0, w0_t, w1_t);
680 make_unicode (w1, w2_t, w3_t);
682 w3_t[2] = out_len * 8 * 2;
689 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
690 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
691 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
692 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
693 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
694 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
695 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
696 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
697 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
698 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
699 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
700 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
701 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
702 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
703 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
704 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
706 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
707 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
708 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
709 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
710 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
711 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
712 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
713 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
714 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
715 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
716 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
717 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
718 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
719 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
720 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
721 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
723 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
724 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
725 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
726 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
727 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
728 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
729 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
730 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
731 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
732 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
733 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
734 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
735 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
736 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
738 if (allx (s2 != ((d + MD4M_D) >> 16))) continue;
740 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
741 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
754 transform_netntlmv1_key (a, b, key);
759 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
763 _des_crypt_encrypt (iv1, data, Kc, Kd, s_SPtrans);
769 transform_netntlmv1_key (((b >> 24) | (c << 8)), ((c >> 24) | (d << 8)), key);
771 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
775 _des_crypt_encrypt (iv2, data, Kc, Kd, s_SPtrans);
781 const u32x r0 = iv1[0];
782 const u32x r1 = iv1[1];
783 const u32x r2 = iv2[0];
784 const u32x r3 = iv2[1];
786 #include VECT_COMPARE_M
790 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_m08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
794 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_m16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
798 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_s04 (__global pw_t *pws, __global gpu_rule_t * rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 rules_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
804 const u32 lid = get_local_id (0);
810 const u32 gid = get_global_id (0);
814 pw_buf0[0] = pws[gid].i[ 0];
815 pw_buf0[1] = pws[gid].i[ 1];
816 pw_buf0[2] = pws[gid].i[ 2];
817 pw_buf0[3] = pws[gid].i[ 3];
821 pw_buf1[0] = pws[gid].i[ 4];
822 pw_buf1[1] = pws[gid].i[ 5];
823 pw_buf1[2] = pws[gid].i[ 6];
824 pw_buf1[3] = pws[gid].i[ 7];
826 const u32 pw_len = pws[gid].pw_len;
832 __local u32 s_SPtrans[8][64];
833 __local u32 s_skb[8][64];
835 s_SPtrans[0][lid] = c_SPtrans[0][lid];
836 s_SPtrans[1][lid] = c_SPtrans[1][lid];
837 s_SPtrans[2][lid] = c_SPtrans[2][lid];
838 s_SPtrans[3][lid] = c_SPtrans[3][lid];
839 s_SPtrans[4][lid] = c_SPtrans[4][lid];
840 s_SPtrans[5][lid] = c_SPtrans[5][lid];
841 s_SPtrans[6][lid] = c_SPtrans[6][lid];
842 s_SPtrans[7][lid] = c_SPtrans[7][lid];
844 s_skb[0][lid] = c_skb[0][lid];
845 s_skb[1][lid] = c_skb[1][lid];
846 s_skb[2][lid] = c_skb[2][lid];
847 s_skb[3][lid] = c_skb[3][lid];
848 s_skb[4][lid] = c_skb[4][lid];
849 s_skb[5][lid] = c_skb[5][lid];
850 s_skb[6][lid] = c_skb[6][lid];
851 s_skb[7][lid] = c_skb[7][lid];
853 barrier (CLK_LOCAL_MEM_FENCE);
855 if (gid >= gid_max) return;
861 const u32 s0 = salt_bufs[salt_pos].salt_buf[0];
862 const u32 s1 = salt_bufs[salt_pos].salt_buf[1];
863 const u32 s2 = salt_bufs[salt_pos].salt_buf[2];
874 const u32 search[4] =
876 digests_buf[digests_offset].digest_buf[DGST_R0],
877 digests_buf[digests_offset].digest_buf[DGST_R1],
878 digests_buf[digests_offset].digest_buf[DGST_R2],
879 digests_buf[digests_offset].digest_buf[DGST_R3]
886 for (u32 il_pos = 0; il_pos < rules_cnt; il_pos++)
916 const u32 out_len = apply_rules (rules_buf[il_pos].cmds, w0, w1, pw_len);
918 append_0x80_2 (w0, w1, out_len);
925 make_unicode (w0, w0_t, w1_t);
926 make_unicode (w1, w2_t, w3_t);
928 w3_t[2] = out_len * 8 * 2;
935 MD4_STEP (MD4_Fo, a, b, c, d, w0_t[0], MD4C00, MD4S00);
936 MD4_STEP (MD4_Fo, d, a, b, c, w0_t[1], MD4C00, MD4S01);
937 MD4_STEP (MD4_Fo, c, d, a, b, w0_t[2], MD4C00, MD4S02);
938 MD4_STEP (MD4_Fo, b, c, d, a, w0_t[3], MD4C00, MD4S03);
939 MD4_STEP (MD4_Fo, a, b, c, d, w1_t[0], MD4C00, MD4S00);
940 MD4_STEP (MD4_Fo, d, a, b, c, w1_t[1], MD4C00, MD4S01);
941 MD4_STEP (MD4_Fo, c, d, a, b, w1_t[2], MD4C00, MD4S02);
942 MD4_STEP (MD4_Fo, b, c, d, a, w1_t[3], MD4C00, MD4S03);
943 MD4_STEP (MD4_Fo, a, b, c, d, w2_t[0], MD4C00, MD4S00);
944 MD4_STEP (MD4_Fo, d, a, b, c, w2_t[1], MD4C00, MD4S01);
945 MD4_STEP (MD4_Fo, c, d, a, b, w2_t[2], MD4C00, MD4S02);
946 MD4_STEP (MD4_Fo, b, c, d, a, w2_t[3], MD4C00, MD4S03);
947 MD4_STEP (MD4_Fo, a, b, c, d, w3_t[0], MD4C00, MD4S00);
948 MD4_STEP (MD4_Fo, d, a, b, c, w3_t[1], MD4C00, MD4S01);
949 MD4_STEP (MD4_Fo, c, d, a, b, w3_t[2], MD4C00, MD4S02);
950 MD4_STEP (MD4_Fo, b, c, d, a, w3_t[3], MD4C00, MD4S03);
952 MD4_STEP (MD4_Go, a, b, c, d, w0_t[0], MD4C01, MD4S10);
953 MD4_STEP (MD4_Go, d, a, b, c, w1_t[0], MD4C01, MD4S11);
954 MD4_STEP (MD4_Go, c, d, a, b, w2_t[0], MD4C01, MD4S12);
955 MD4_STEP (MD4_Go, b, c, d, a, w3_t[0], MD4C01, MD4S13);
956 MD4_STEP (MD4_Go, a, b, c, d, w0_t[1], MD4C01, MD4S10);
957 MD4_STEP (MD4_Go, d, a, b, c, w1_t[1], MD4C01, MD4S11);
958 MD4_STEP (MD4_Go, c, d, a, b, w2_t[1], MD4C01, MD4S12);
959 MD4_STEP (MD4_Go, b, c, d, a, w3_t[1], MD4C01, MD4S13);
960 MD4_STEP (MD4_Go, a, b, c, d, w0_t[2], MD4C01, MD4S10);
961 MD4_STEP (MD4_Go, d, a, b, c, w1_t[2], MD4C01, MD4S11);
962 MD4_STEP (MD4_Go, c, d, a, b, w2_t[2], MD4C01, MD4S12);
963 MD4_STEP (MD4_Go, b, c, d, a, w3_t[2], MD4C01, MD4S13);
964 MD4_STEP (MD4_Go, a, b, c, d, w0_t[3], MD4C01, MD4S10);
965 MD4_STEP (MD4_Go, d, a, b, c, w1_t[3], MD4C01, MD4S11);
966 MD4_STEP (MD4_Go, c, d, a, b, w2_t[3], MD4C01, MD4S12);
967 MD4_STEP (MD4_Go, b, c, d, a, w3_t[3], MD4C01, MD4S13);
969 MD4_STEP (MD4_H , a, b, c, d, w0_t[0], MD4C02, MD4S20);
970 MD4_STEP (MD4_H , d, a, b, c, w2_t[0], MD4C02, MD4S21);
971 MD4_STEP (MD4_H , c, d, a, b, w1_t[0], MD4C02, MD4S22);
972 MD4_STEP (MD4_H , b, c, d, a, w3_t[0], MD4C02, MD4S23);
973 MD4_STEP (MD4_H , a, b, c, d, w0_t[2], MD4C02, MD4S20);
974 MD4_STEP (MD4_H , d, a, b, c, w2_t[2], MD4C02, MD4S21);
975 MD4_STEP (MD4_H , c, d, a, b, w1_t[2], MD4C02, MD4S22);
976 MD4_STEP (MD4_H , b, c, d, a, w3_t[2], MD4C02, MD4S23);
977 MD4_STEP (MD4_H , a, b, c, d, w0_t[1], MD4C02, MD4S20);
978 MD4_STEP (MD4_H , d, a, b, c, w2_t[1], MD4C02, MD4S21);
979 MD4_STEP (MD4_H , c, d, a, b, w1_t[1], MD4C02, MD4S22);
980 MD4_STEP (MD4_H , b, c, d, a, w3_t[1], MD4C02, MD4S23);
981 MD4_STEP (MD4_H , a, b, c, d, w0_t[3], MD4C02, MD4S20);
982 MD4_STEP (MD4_H , d, a, b, c, w2_t[3], MD4C02, MD4S21);
984 if (allx (s2 != ((d + MD4M_D) >> 16))) continue;
986 MD4_STEP (MD4_H , c, d, a, b, w1_t[3], MD4C02, MD4S22);
987 MD4_STEP (MD4_H , b, c, d, a, w3_t[3], MD4C02, MD4S23);
1000 transform_netntlmv1_key (a, b, key);
1005 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
1009 _des_crypt_encrypt (iv1, data, Kc, Kd, s_SPtrans);
1016 transform_netntlmv1_key (((b >> 24) | (c << 8)), ((c >> 24) | (d << 8)), key);
1018 _des_crypt_keysetup (key[0], key[1], Kc, Kd, s_skb);
1022 _des_crypt_encrypt (iv2, data, Kc, Kd, s_SPtrans);
1034 const u32x r0 = iv1[0];
1035 const u32x r1 = iv1[1];
1036 const u32x r2 = iv2[0];
1037 const u32x r3 = iv2[1];
1039 #include VECT_COMPARE_S
1043 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_s08 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)
1047 __kernel void __attribute__((reqd_work_group_size (64, 1, 1))) m05500_s16 (__global pw_t *pws, __global gpu_rule_t *rules_buf, __global comb_t *combs_buf, __global bf_t *bfs_buf, __global void *tmps, __global void *hooks, __global u32 *bitmaps_buf_s1_a, __global u32 *bitmaps_buf_s1_b, __global u32 *bitmaps_buf_s1_c, __global u32 *bitmaps_buf_s1_d, __global u32 *bitmaps_buf_s2_a, __global u32 *bitmaps_buf_s2_b, __global u32 *bitmaps_buf_s2_c, __global u32 *bitmaps_buf_s2_d, __global plain_t *plains_buf, __global digest_t *digests_buf, __global u32 *hashes_shown, __global salt_t *salt_bufs, __global void *esalt_bufs, __global u32 *d_return_buf, __global u32 *d_scryptV_buf, const u32 bitmap_mask, const u32 bitmap_shift1, const u32 bitmap_shift2, const u32 salt_pos, const u32 loop_pos, const u32 loop_cnt, const u32 combs_cnt, const u32 digests_cnt, const u32 digests_offset, const u32 combs_mode, const u32 gid_max)